Google Research releases TabFM, a zero-shot foundation model for tabular data

Google Research published TabFM on June 30, 2026, a foundation model that performs zero-shot classification and regression on tabular data – spreadsheets and database tables – with no dataset-specific training, tuning or feature engineering. It reads a full table as a single prompt and predicts through in-context learning.

The pitch is to cut the per-dataset engineering that tabular machine learning usually demands. Where a data team would normally build and tune a model for each table, TabFM is meant to run against a new one directly, the way a language model answers a new prompt without retraining.

The architecture combines alternating row and column attention, in the style of the TabPFN approach, with row compression and an in-context-learning Transformer drawn from the TabICL method. Google trained it entirely on hundreds of millions of synthetic datasets generated from structural causal models, a choice the lab attributes to the scarcity of diverse open tabular datasets. Google Research evaluated TabFM and an ensemble variant, TabFM-Ensemble, on the TabArena benchmark, spanning 38 classification and 13 regression datasets of 700 to 150,000 samples.

The model and code are available on GitHub and Hugging Face now, with BigQuery SQL integration via an AI.PREDICT function planned.

The benchmark results are Google’s own and not yet independently reproduced, and training solely on synthetic data raises the standard question of whether performance holds on the messy, biased real-world tables practitioners actually load. TabArena’s datasets top out at 150,000 samples, leaving open how the model behaves on the far larger tables common in production. Whether outside teams reproduce the results on their own data is the test that matters.